Description

Sprint-2 adds classical NLP baselines for the Emotion/State notes dataset.

What’s included

• New script: AI Guardian/emotional_baseline.py
• Artifacts (in AI Guardian/out_final/):
  • sprint2_report.md (summary & metrics)
  • sprint2_results.csv (single split), sprint2_cv_results.csv, sprint2_cv_summary.csv
  • confusion_matrix.png
  • feature_preview.txt (evidence of no label tokens in vocab)
  • settings.json (repro config)

Methods

• Features: BoW(1), TF-IDF(1–2), TF-IDF(1–3) + optional lexicon counts
• Models: Logistic Regression & Linear SVM (class_weight="balanced")
• Leakage guard: removes label tokens (normal/sick/uncomfortable + variants) via custom preprocessor + stopword list, with a tripwire that fails if any leak into the vocab
• Group-aware split (GroupShuffleSplit) + 5-fold GroupKFold CV; dedup after scrubbing

Headline results (use CV as primary)

• Best CV: TF-IDF(1–2) + LinearSVM ≈ 0.98 accuracy, 0.956 ± 0.032 macro-F1
• Single split shown in confusion_matrix.png (informational)

Todos

• Tested and working locally (sprint2nlp conda env)
• Code follows project style (black-ish formatting, no notebooks committed for pipeline)
• Self-reviewed; checked feature preview for label leakage
• Documentation: sprint2_report.md + settings.json
• Request review from 2 devs (ML + app)

How to test

1. Ensure conda env has sklearn/pandas/numpy/scipy/matplotlib.

   conda activate sprint2nlp